Noise-Contrastive Estimation Based on Relative Neighbour Sampling for Unsupervised Image Embedding Learning

Many unsupervised learning algorithms have been proposed to avoid the inconvenience of data labeling. The nature of neural networks is also gradually being explored by unsupervised learning. In this paper, we focus on the sampling strategy for unsupervised images embedding learning via instance discrimination. A new sampling method is proposed based on the observation that different samples contribute unequally to training, which pays more attention to the neighbours. The proposed sampling method is beneficial and efficient for images embedding learning. The results on benchmark data show that the proposed sampling method is robust and outperforms the compared method.